(j3.2006) Fwd: MPI Forum aftermath

[Craig Rasmussen]

I attended an MPI Forum workshop last week where I'm heading a  
committee to come up with new MPI Fortran bindings based on C interop  
(with Aleks' help).  There I talked with a graduate student at  
Indiana who is doing research on collective operations.  Since  
collectives have been pulled out of 2008 and into a TR, I though it  
might be appropriate if Torsten could attend the next J3 meeting and  
give a presentation on his work (discovering collective usage in MPI  
applications) and on collective proposals for MPI 3.0.

Should I invite him to give a presentation, perhaps on Monday May 12?

Regards,
Craig



Begin forwarded message:

> From: Torsten Hoefler <htor at cs.indiana.edu>
> Date: March 13, 2008 10:25:34 PM MDT
> To: crasmussen at lanl.gov
> Subject: MPI Forum aftermath
>
> Hello,
> we talked at the MPI Forum about several Fortran issues and Co Array
> Fortran.
>
> The function call to prevent the compiler from moving the MPI_Wait()
> after accessing the communicated memory is MPI_Get_address()
> (http://mpi.deino.net/mpi_functions/MPI_Get_address.html).
>
> I also said that I would have to check some conference dates in May to
> see if I can follow your invitation. It looks good. I will be in
> Bloomington between May 5th and 17th and I would really like to  
> come and
> talk to the CAF group. I can give a talk about the newly proposed
> collectives as well as the established ones. Non-blocking collectives
> and latency-hiding/overlap is a hot topic (especially as compilation
> targets, cf. UPC "Optimizing Bandwidth Limited Problems Using One- 
> Sided
> Communication and Overlap" and "Communication Optimizations for
> Fine-grained UPC Applications").
>
> I am not involved in UPC, but my research with non-blocking  
> collectives
> and overlap (cf. "Implementation and Performance Analysis of
> Non-Blocking Collective Operations for MPI", "Optimizing a Conjugate
> Gradient Solver with Non-Blocking Collective Operations" and  
> "Optimizing
> non-blocking Collective Operations for InfiniBand") seems to lead
> towards compiled communication (it seems very complicated to do
> manually). Looking at current widely used applications (I investigated
> MILC, MPQC, POP, ABINIT and TDDFT/octopus, all written in Fortran  
> 77/90
> so far) looks like collectives and neighbor exchange (i.e., next
> generation collectives) are the most important operations and
> non-blocking collectives could have some potential. I am highly
> interested in the compiler-based approaches.
>
> Please let me know if you have additional applications (I would need
> sources and input files) to investigate.
>
> Thanks & Best,
>   Torsten
>
> -- 
> Indiana University    | http://www.indiana.edu
> Open Systems Lab      | http://osl.iu.edu/
> 150 S. Woodlawn Ave.  | Bloomington, IN, 474045-7104 | USA
> Lindley Hall Room 135 | +01 (812) 855-3608

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://j3-fortran.org/pipermail/j3/attachments/20080318/34852e55/attachment.html